Clinical Decision Support Systems

Clinical Decision Support



Clinical decision support (CDS) is going to be an important element of Meaningful Use Stages 2 and 3. Decision support functions will go far beyond those of drug allergy alerts and drug interaction checking that are necessary in Stage 1. In this post I will discuss diagnostic CDS. Wikipedia has an excellent general introductory discussion of CDS. You might also want to check the links to Isabel and SimulConsult. These are two examples of diagnostic CDS systems that are already commercially available. The former uses subject matter experts to generate content and the latter utilizes a computational wiki.


There have been several times in my career where diagnostic CDS would have been helpful. First, one must understand how doctors establish diagnoses. Usually this is done via a series of heuristics, rules of thumb, that often work very efficiently for the diagnosis of common conditions. The problem with heuristics is that there are a number of biases that can lead one in the wrong direction and thus to the wrong diagnosis. If the diagnosis for a patient is not correct, then the treatment instituted is likely to be incorrect. Let me provide a few examples from my clinical experience.


Years ago, I evaluated a patient in the ortho clinic with wrist pain. His usual clinician was away so I was seeing the patient. He had been treated for several months with non-steroidal anti-inflammatories with minimal relief of symptoms. Overuse syndromes, minor sprains, and arthritis are common causes of joint pain so the working diagnosis was one of these. I reviewed his previous x-rays and was struck by the marked washing out of mineral (a finding known as osteopenia) from all the wrist bones. I recalled that among the possible etiologies of this x-ray finding is tuberculosis. It also turned out that the patient had a racial background that showed a susceptibility to tuberculosis infection. I performed a synovial biopsy of the wrist that came back positive for TB at 4 weeks. The treatment was changed from NSAIDs to triple anti-tuberculous antibiotic therapy. The initial treating physician thought the problem a common one. Here availability bias resulted in a diagnostic error. Now I would like to relate another example.


I was called to the operating room to see a patient with an infected foot. The treating surgeon was a very good orthopedists. He was taking a reasonable approach by treating the presumed infection/ulcer with surgery and antibiotics. The patient had what appeared to be a large ulcer with pus at his heel/instep area. Initially, there did not appear to be a good reason for this patient to have a spontaneous foot infection. As a consultant, I carefully reviewed the chart. Two data elements that initially were not given proper weight led me to the correct diagnosis. The patient had a history of inflammatory bowel disease. In fact, he had a small fistula from bowel to his anterior abdominal wall (this is characteristic of regional enteritis, one of the inflammatory bowel diseases.) Also, the very compulsive ward nurses took a photograph of the initial appearance of the foot. That photo showed a bullous lesion (a clear blister) rather than a pustule (abscess.) Only a few conditions cause bullae. Since I had an interest in dermatology as a medical student I knew where to look for the diagnostic label. Pyoderma gangranosum is a skin condition associated with inflammatory bowel disease that starts as a bullous lesion. The diagnosis is made by clinical history, appearance of the lesion, and microscopy of a biopsy of the lesion. The treatment, surprisingly, is high dose intravenous steroids, something we would never consider in the treatment of infections because steroids inhibit the body’s immune defenses. The healing process was slow once the treatment was changed but the patient did not require further surgery. The heuristic bias here was representativeness.


It is possible that both of these incorrect diagnoses could have been avoided if the treating physicians had had EHRs with diagnostic CDS running in the background. Then the diagnosis in the first case would have keyed on the differential of diffuse periarticular osteopenia- either the radiologist or the orthopedist would have needed to enter the proper term in the record. In the second case, an entry of a bullous skin lesion on the physical exam or a skin biopsy narrative would have been needed to feed the CDS system with the necessary data to help with the correct diagnosis. This emphasizes the point that use of CDS is not cookbook medicine. To have good CDS you have to have expert clinicians who can provide the terms from history, physical exam, laboratory and x-ray studies that when coded properly will allow the CDS system to do its job. Otherwise we have the problem of garbage-in, garbage-out. A recent post on John Moehrke's blog discusses some of the data gathering issues with CDS systems as well as some relevant privacy concerns.


One of the problem with heuristics is that the probabilities are not really what the clinician thinks. Bayes' theorem is a mathematical formula that addresses probability of diagnoses before and after tests. This is familiar to many who work in the CDS field but not often considered or used by others. Clinicians are generally not well trained in probability theory as it relates to diagnosis and interpretation of tests. Other important statistical terms that should be considered in clinical diagnosis are sensitivity, specificity, positive predictive value, negative predictive value, and the accuracy of given tests. One also needs to know which of these statistical measures regarding tests are most important to weigh when “ruling in” a condition or in “ruling out” a condition. This tragically hit too close to home recently when my brother-in-law died suddenly from a pulmonary embolism.


Early one week he had sudden onset of sharp right-sided chest/abdominal/flank pain. He was seen by a physician's assistant at an urgent care center. He was discharged with a recommendation to obtain an abdominal ultrasound on an elective basis. Several days later he got on an airplane and traveled from the west coast to the midwest. The evening of arrival he experienced severe right sided flank pain, shortness of breath, low grade fever, low blood oxygen content, and hemoptysis. The differential included pneumonia and pulmonary embolism. The history, physical, and lab tests should have led to the conclusion based on probabilities that pulmonary embolism was just as likely as pneumonia. When a CT angiogram of the chest did not show evidence of an embolism then the clinicians wrongly concluded that he had pneumonia. The bias here was anchoring. Treatment was with antibiotics and pulmonary support. Further studies should have been done that most likely would have revealed the correct diagnosis. Several days after discharge from the hospital and another cross-country airplane trip, my brother-in-law collapsed at home and was dead 2 hours later. Postmortem showed no signs of pneumonia but rather acute and subacute pulmonary embolism. This devastated my family. I cannot help but wonder if a good CDS system had been used that the clinicians would have been reminded to do the correct tests and then institute the correct treatment with anticoagulants (blood thinners) rather than antibiotics. This is exactly what diagnostic CDS is all about. It would have saved his life.